fix: record unanswered questions an unattended turn could not ask

[ZuyiZhou]

Summary

An unattended turn could ask a question and leave no trace a person can read.

raven agent -m already lists calls the permission gate refused (#708). It said nothing about a question nobody could answer. On the web, a sub-agent turn with no page subscribed to the session binds no asker (start_ask_turn(None) in raven/rpc/spine.py). The host ask_user tool still tells the model "user did not answer; proceed with best judgment". That sentence is the tool result the model reads. Opening the session later showed no event for the question.

Questions that were never put to anyone are now recorded on the turn: a sub-agent ask_user or elicitation with no asker, a question that gives up waiting for the conversation lock, a round trip that is structurally unavailable, and a host ask_user call whose answer comes back empty. A one-shot call has no broker at all, so it returns at that guard; the questions _prepare accepts are recorded there too, and the same error string is what the model still reads. A call _prepare rejects is not recorded. The turn's save writes one assistant entry with a question_unanswered notice, which is what a reopened session draws. A one-shot prints the same questions after the reply. That list does not change the exit status: the turn continued, which is what the model was told to do. A refusal is still the run that exits 3.

Type

• Fix
• Feature
• Docs
• CI / tooling
• Refactor
• Other

Verification

• Relevant tests pass locally
• Relevant lint / type checks pass locally
• User-facing docs or screenshots are updated when needed

uv run --frozen --python 3.12 --extra dev ruff check raven/agent/loop/turn_path.py raven/agent/tools/ask_user.py tests/test_ask_user_tool.py tests/test_cli_one_shot_spine.py tests/test_agents_research_flow_report_shape.py
  All checks passed

uv run --frozen --python 3.12 --extra dev ruff format --check <same files>
  already formatted

uv run --frozen --python 3.12 --extra dev pytest -q \
  tests/test_ask_user_tool.py::test_a_call_with_no_broker_records_the_normalized_questions \
  tests/test_ask_user_tool.py::test_error_paths_return_plain_strings \
  tests/test_cli_one_shot_spine.py::test_a_one_shot_ask_user_call_with_no_broker_is_what_gets_reported \
  tests/test_agents_research_flow_report_shape.py::test_the_bounce_scaffolding_never_becomes_history \
  tests/test_agent_loop_session_stamps.py::test_an_unanswered_question_is_stored_as_a_notice
  5 passed

cd ui-tui && npm run lint:i18n
  i18n: generated catalogue is up to date

PYTHONPATH=. uv run --frozen --python 3.12 --extra dev python scripts/check_source_language.py origin/main...HEAD
  exit 0

make check-commits && PR_TITLE="fix: record unanswered questions an unattended turn could not ask" make check-pr-title
  exit 0

git diff --check origin/main...HEAD
  exit 0

git merge-tree --write-tree HEAD origin/main
  clean

Reverting raven/agent/tools/ask_user.py to its parent makes both new no-broker tests fail with an empty record. Reverting raven/agent/loop/turn_path.py to the parent, where filing was a method, makes test_the_bounce_scaffolding_never_becomes_history raise AttributeError: '_Loop' object has no attribute '_file_unanswered'. Both files were restored afterwards. The earlier reverts (notice not stored, ACP ask and elicitation not recorded, one-shot report not printed) still stand for those paths.

The full suite was not run. A run of the touched CLI modules in this terminal also failed existing assertions that search Rich output for a plain substring (help flags, and 1 action(s) were refused). Those assertions fail the same way with this diff's refusal printer left unchanged. The type checker was not rerun locally; CI's python lint job runs it on the head.

Risk

• Security impact considered
• Backward compatibility considered
• Rollback path is clear for risky changes

The model is still told to continue. Nothing new is approved. The one-shot exit status is unchanged: unanswered questions are reported and do not by themselves exit 3. The notice is one more assistant entry on the session, with notice.kind question_unanswered, so a client that already draws runtime notices draws this one. Rollback is reverting the commit.

Related Issues

#708

[gloryfromca]

Blocking: the no-broker one-shot path must record its unanswered questions before this can merge.

I left one inline finding. I reviewed the actual PR delta against GitHub's current base (2f5027cdf); the local github/main ref was stale, so I also separated the two already-merged intervening commits from this PR. I covered the repository rules and runtime/TUI domain guidance, the diff, production callers and persistence paths, relevant history, backward compatibility, test quality (including whether the new tests bypass production behavior), and the runtime architecture boundary.

Verification:

• uv run pytest tests/test_acp_ask_user.py tests/test_acp_elicitation.py tests/test_agent_loop_session_stamps.py tests/test_ask_user_tool.py tests/test_cli_agent_commands.py tests/test_cli_one_shot_spine.py -x - 235 passed.
• uv run python scripts/check_source_language.py 2f5027cdfc279a96ba3fb69ac351f5652b63f0dc...HEAD - passed.
• git diff --check 2f5027cdfc279a96ba3fb69ac351f5652b63f0dc...HEAD - passed.
• A direct no-broker AskUserTool reproduction returned the existing configuration error and left turn.unanswered empty.

make check-source-language could not run because make is unavailable in this environment; the canonical checker it wraps passed when invoked directly.

[gloryfromca]

No blockers; this can merge as far as I am concerned.

The prior blocker is fixed: the production no-broker and missing-conversation paths now record normalized, valid questions before preserving their existing error returns, and the one-shot regression test exercises that path through AskUserTool and the runner. Moving unanswered-notice filing to a module-level helper also preserves callers that borrow _save_turn on a stub.

For this revision I reviewed the delta from 73725bb71050, the updated full PR diff, the affected callers and persistence path, compatibility of existing error returns, the regression tests (including whether they reach production behavior), repository rules/domain terminology, and the existing runtime architecture boundary. I found no additional issue.

Verification:

• uv run pytest tests/test_acp_ask_user.py tests/test_acp_elicitation.py tests/test_agent_loop_session_stamps.py tests/test_ask_user_tool.py tests/test_cli_agent_commands.py tests/test_cli_one_shot_spine.py tests/test_agents_research_flow_report_shape.py -x - 283 passed.
• uv run python scripts/check_source_language.py 2f5027cdfc279a96ba3fb69ac351f5652b63f0dc...HEAD - passed.
• git diff --check 2f5027cdfc279a96ba3fb69ac351f5652b63f0dc...HEAD - passed.